Exploring Fast Style Transfer

We investigate enhancements to fast neural style transfer networks by including a perceptual loss term for a pre-set ”object” image (henceforth referred to as ”object loss”), in addition to the existing methods of using perceptual style losses and perceptual content losses. We find that optimizing the transfer network for object loss alone (without style loss) does not produce satisfactory results as the network is unable to learn to transform the input image into one whose perceptual loss features sophisticatedly match that of the reference object image. However, by including the style loss together with the object loss, we find that the object loss is able to change the type of style and content that is applied to the input image. We also investigate visualizations of style transfer networks via feedback, which shows the patterns learnt by the network from the style image, independent of a content image. Figure 1: Style Transfer Example